previous, up, next

Background

In many existing neural network models, e.g. multi-layer-perceptrons, radial basis function networks or self-organizing feature maps the topology of the network has to be chosen in advance (before training). This can be a difficult step since neural networks are often used for problems, where the statistical properties of the training data are not well known. Choosing a too large network may result in overfitting the training data and poor generalization to new data. A too small network may not be able to capture the relations underlying the training data and will show poor results on new data, too.

For this reason, methods to adapt the network topology (or in general to change the number of free parameters) have been investigated extensively. Examples include Optimal Brain Damage (Le Cun et al., 1990), Cascade Correlation (Fahlman & Lebiere, 1990), the upstart algorithm (Frean, 1990), weight decay (Krogh et al, 1992) and weight elimination (Weigend, Rumelhart and Huberman,1991).

(Last updated: Feb. 7, 1997)